-
Notifications
You must be signed in to change notification settings - Fork 2k
Description
Describe the bug
When using the OpenPose ControlNet model (control_v11p_sd15_openpose_fp16.safetensors or OpenPoseXL2.safetensors) with Preprocessor: none and a manually-created skeleton from the OpenPose Editor, the generated images do not follow the provided pose.
This happens consistently, especially with arm positions being ignored. I've tested both SDXL (juggernautXL) and SD 1.5 (chilloutmix) checkpoints. The issue persists even with Control Weight set to 1.0–1.3 and Control Mode set to "ControlNet is more important."
To Reproduce
Steps to reproduce the behavior:
- Use A1111 on macOS with MPS backend
- Load an SD 1.5 or SDXL model
- Create a pose in the OpenPose Editor and send it to
txt2img - Enable ControlNet:
- Preprocessor: none
- Model:
control_v11p_sd15_openpose_fp16orOpenPoseXL2 - Control Mode: ControlNet is more important
- Control Weight: 1.0+
- Prompt: descriptive body pose (e.g., arms stretched forward)
- Generate
Expected behavior
The output should follow the full-body pose defined in OpenPose Editor, especially for clear limb placements.
Actual behavior
The body pose is ignored partially or fully, especially the arms. Sometimes only the legs follow. This happens on both SDXL and SD 1.5 pipelines.
Environment
- Hardware: Mac Studio M2 Ultra
- Platform: macOS (MPS backend)
- WebUI: Automatic1111 (latest)
- ControlNet version: v1.1.455
- Python: 3.10.x
- Torch: 2.6.0
- Diffusers: 0.31.0
- Transformers: 4.30.2
- Accelerate: 0.21.0
Additional context
- Same pose input works with IP-Adapter
plus-facefor identity (partially), but IP-Adapter cannot enforce body pose. - Preprocessor set to
none, pose built entirely in OpenPose Editor - Confirmed OpenPose models load correctly (
control_v11p_sd15_openpose_fp16.safetensorsandOpenPoseXL2.safetensors) - Confirmed pose preview appears correctly in ControlNet image panel